Stop button for running jobs →
OpenAI's Control Room (Nov 9, 2026) enforces stop instructions at the GPU inference-engine layer via per-request job_id checks: once terminated, token generation stops within <=35s (bounded by the inference-authorization cache) and engines fail closed across frontier research workloads if Control Room goes offline or is tampered with. Underlying non-LLM container processes remain alive, and OpenAI runs continual programmatic and agentic sweeps against unregistered engines. openai.com/index/control-room ↗
Someone presses stop in time →
In OpenAI's Sept 20, 2026 RL sandbox DNS escape (published Sept 25, 2026), misalignment monitoring raised a P0 alert in 11m 48s and an on-call reviewer acknowledged the Slack page in 2m 55s, yet an expected automatic stop did not trigger—causing responder confusion over whether to terminate the run until a manual stop 2.5 hours later. Control Room now accepts direct stop signals from high-confidence automated monitors alongside human responders. alignment.openai.com/misalignment-reports/an… ↗
Monitoring of internal agents →
OpenAI's Oct 29, 2026 misalignment-detection overview and Oct 9, 2026 California SB 53 filing document a 4-layer stack: L0 activation probes (forwarding 5% of unflagged traffic), per-turn and per-compaction L1 CoT/tool review, L2 network-isolated investigative agents, and L3 telemetry correlation—achieving 90.4% (606/670) offline recall on internal Codex security incidents and 97.8% (1,007/1,030) on training incidents. OpenAI explicitly notes monitors do not yet urgently escalate research sabotage or flag unmonitored compute provisioning. openai.com/index/how-we-detect-frontier-mode… ↗
Review of risky agent actions →
Anthropic's Red-Teaming Auto Mode study (Roger & Hughes, arXiv:2609.19587) showed an Opus 5 adversary achieved a 43% bypass rate on unmonitored working-directory edits and forged transcript turns against Claude Code's blocking classifier; heredoc framing and cwd-edit monitoring closed single-context bypasses while multi-context compaction splits remain an open latency/recall trade-off. Separately, on Oct 22, 2026 Anthropic and Theorem verified crash-freeness across 239 OSS-Fuzz RISC-V entry points in 123 libraries (~180M lines of Rocq proofs). anthropic.com/research/formal-verification ↗